aws-deepracer-community / deepracer-log-guru Goto Github PK

I feel like it is slower than it used to be - not so slick - need to see if any of the stats population is rather slow (if so - can we move those to parallel thread etc... ? )

Better than reward20 reward100 and so on

FileNotFoundError: [Errno 2] No such file or directory: '/Users/xxxxxxxx/Downloads/test-aws-clone/logs/training\training-20210518053907-55l2utpgSgqiuiLkaMGl

Show the car as a simple rectangle - nothing special
Chosen episodes only (with some limit e.g. 10)
Each one a different colour
Presumably all from same starting point?

Or consider at least being able to customize certain aspects via extra settings in the personal_configuration file

Two obvious issues are (a) white flashes going from track to graph analysis (b) slight changes in the size of the right hand frame on different analysis options

One user has set wraplength to 1000 which helps apparently

Idea is to search all waypoints or short sections to see if certain actions, positions, speed or bearing have a clear correlation with either progress % or overall lap time

An incomplete iteration can produce a large spike or dip at the end of a graph depending on how uneven performance is across various starting points - note for example that the AWS Console only adds results for completed iterations (although it does add incomplete evaluation results - which can be a bit spikey)

The biggest issue (for me) is that this can mean the graph scale is reduced because a lot of vertical space is given to the spike.

Useful especially for lap time (including lap prediction) views - the scale and correlations can be messed up by an episode that spun

Various internal matplotlib related warnings. Don't try to plot missing or limited amount of data etc.

Quarters is proving a nice way to abstract training progress above the basic reward graph in trend is sometimes hard to see if there's a lot of seesawing

Also can show things like range of completed lap times etc. & lots more besides to see which things model is getting better at, and which are getting worse .... (and in mini graphs so without the hassle of clicking radio buttons to flick between them, unlike the bigger analysis views)

I mean ability to extract the corresponding reward function inputs for position i.e. waypoint id, distance from centre and left/right indicator which would then give users a VERY easy way to build a new reward function based on that route ...

Basically this means we don't get "lost" between views - for example, on speed analysis, the event dot can be highlighted, and also the same arrow keys can be used to navigate backwards or forwards along the actions/steps ...

Switching track effectively means no log is loaded ... menu should probably revert to the limited menu on startup (fewer options) until user has opened a log file for the newly chosen track

User reported things like "filtered" and also especially the different colour schemes for the episode route analysis

Ideally record in the meta too for next time file is loaded?

Needs more thought, but seems like powerful way to tag episodes if during analysis you find one you want to come to (e.g. because it is your best "route" around a track that you want to keep)

Something like click on the plotted dot and it takes you to route analysis?

Rather than having to use << and >>

Sometimes it is irritating, nice to turn it off via a checkbox

Useful for users to provide a familiar point of reference
Useful for testing to prove our data analysis is not completely wrong - i.e. our graph should look identical to the console, doh!

Support classic drag-box selection of zoom area using mouse ...

Fix the lap time predictor especially to focus on predictions rather than its previous remit

Particular issue with usability is that checking a different button takes time to refresh - really the button blob should move as soon as the user clicks - and the screen can then continue to refresh in the next few seconds ... ?

Also should optimize the convergence calculations themselves using a mixture of:
(a) cache previous results (i.e. fast switching too and fro)
(b) use multi-threading [ now raised as issue #65 ]
(c) perhaps ndarray will be faster for building the data (i.e. faster indexing of the large array etc.)?

Important pre-requisite before I can integrate with alternative reward functions (see separate issue)

Effect is really bad with debug output, but even when no debug then it can still happen to much less degree depending on width of data items being displayed

Lap predictor is still really a left-over from a "v1" lap time correlation analysis.

But now that "real" laps have their own correlation view, really the predictor should be modified to be more unique to predictions

(some duplication here versus another issue that says to remove duplication in different analysis views)

Followed the Direction as in Installation Guide and upon running guru.py, I get below error.

File "/Users//deep_racer_guru/src/sequences/sequences.py", line 81, in load
with open(_FILENAME, "r") as infile:
FileNotFoundError: [Errno 2] No such file or directory: 'DRG_sequences.json'

i.e. to basically see what the effect will be of different discount factors like 0.9 0.95 0.99 0.995 0.999 etc.

Would be less confusing, especially once I write some proper documentation

SD video in YouTube is a bit fuzzy for text details. Would be better to make them larger (some users might also prefer it ... ?)

With an export button too (would be nice)

For me, the main thing is to be able to see/review how I tinkered with the hyper parameters (coupled with a reminder of the defaults - so easy to see what changed, and by how much ... )